MoZoo: Generating Realistic Animal Fur and Muscle Animation with Video Diffusion Models

Forum topic · 小凯 · 2026-05-17

Summary

MoZoo is a video diffusion framework that generates high-fidelity animal videos—complete with realistic fur and muscle deformation—directly from coarse 3D meshes, bypassing traditional per-strand hair physics simulation. The paper introduces two key attention designs: Role-Aware RoPE (RAR-RoPE), which keeps motion features of different animal body parts separate yet aligned, and Asymmetric Decoupled Attention, which unidirectionally flows appearance information into motion to prevent identity drift. Training uses MoZoo-Data, a synthetic-to-real pipeline combining rendered videos with perfect ground truth and real animal footage via inverse mapping. Evaluation is supported by MoZooBench, a benchmark of 120 coarse mesh-video pairs. The author notes that traditional fur simulation scales poorly (quadratic cost with hair count), while diffusion models learn the coarse-to-fine mapping directly. Reported results show consistent temporal and structural quality across animal skeletons, though the author cautions they only read the paper text and did not verify the released video demos. A broader reflection: when AI can skip physical simulation and generate directly, where does data-driven imagination fail where physics-driven computation succeeds?

This forum post reviews the paper *MoZoo: Unleashing Video Diffusion power in animal fur and muscle simulation* (arXiv:2605.13857), which proposes generating realistic animal animation directly from a coarse 3D skeleton using video diffusion models.

Why animal animation is so hard

Every strand of fur betrays a fake animal. Fur motion involves microsecond-level changes in orientation, bending, and occlusion under wind, and muscle bulging must exactly match skeleton movement. Human viewers are evolutionarily hyper-sensitive to "slightly wrong" animals, so any error reads as fake. Traditional pipelines (manual modeling + physics simulation + frame-by-frame cleanup) are expensive and slow—think of the tens of millions of dollars behind the 2019 *The Lion King*.

MoZoo's approach

Instead of simulating each hair, MoZoo generates high-quality animal video end-to-end from a rough 3D mesh:

Coarse mesh input: a hairless, texture-less skeleton, even roughly shaped, serves as the starting point.
Multimodal guidance: color, texture, and motion can be specified via text or reference images.
Diffusion generation: the model "imagines" what the animal should look like in high-resolution video.

Two architectural ideas (per the author's reading):

Role-Aware RoPE (RAR-RoPE): different body parts (head, legs, tail) have very different motion characteristics; a clever index remapping keeps these "role" motion streams independent while still aligned.
Asymmetric Decoupled Attention: the latent sequence is split into a reference (appearance) block and a motion block, with strictly one-way information flow from appearance to motion—preventing the animal's identity from drifting as it moves.

The author honestly admits the paper lacks quantitative ablation numbers for these modules, so the effectiveness of the hyperparameter-heavy designs remains to be verified.

Synthetic data first

MoZoo-Data: a synthetic-to-real training pipeline. Rendered videos provide perfect ground truth; real animal footage is incorporated via inverse mapping.
MoZooBench: a benchmark of 120 coarse mesh–video pairs—small in count, but building it (especially with furry high-resolution clips) is substantial work.

Reported results and caveats

The paper reports high-fidelity fur simulation with temporal and structural consistency across animal skeletons and layouts. The author notes: they only read the paper's text and did not verify the GitHub video demos, so claims like "more realistic than *The Lion King* (2019)" cannot be assessed. Still, the direction seems sound: traditional fur physics scales quadratically with hair count, while diffusion models learn the coarse-to-fine mapping and sidestep that bottleneck.

Reflection: does simulation still matter?

When AI jumps from input directly to output, skipping modeling→rigging→animation→simulation→rendering, physicists instinctively object: how do you know the result is right without physics? But human animators never computed Navier-Stokes either—they replaced computation with intuition, an efficient built-in physical approximator. MoZoo doesn't replace physics simulation so much as "imagine" animals in a more human way, grounded in statistical learning from real data. The open question: at what boundary does data-driven imagination fail where physics-driven computation succeeds? The paper doesn't answer it—and neither does the author.

Paper information

Title: MoZoo: Unleashing Video Diffusion power in animal fur and muscle simulation
Authors: Dongxia Liu, Jie Ma, Xiaochen Yang, Jiancheng Zhang, Bin Xia, Zhehan Kan, Nisha Huang, Jun Liang, Wenming Yang, Jin Li
arXiv: arXiv:2605.13857
Date: April 8, 2026
Contributions: end-to-end video diffusion generation of fur and muscle animation from coarse meshes; RAR-RoPE and Asymmetric Decoupled Attention; MoZoo-Data synthetic pipeline; MoZooBench benchmark

References

1. Liu, D., et al. (2026). MoZoo: Unleashing Video Diffusion power in animal fur and muscle simulation. arXiv:2605.13857. 2. Ho, J., et al. (2022). Video Diffusion Models. NeurIPS 2022. 3. Singer, U., et al. (2023). Make-A-Video: Text-to-Video Generation without Text-Video Data. ICLR 2023.